Please use this identifier to cite or link to this item:
Title: Homogeneous vector capsules and their application to sufficient and complete data
Authors: Byerly, Adam D.
Advisors: Kalganova, T
Dolins, S
Keywords: Convolutional neural networks;Capsule networks;Dimensionality reduction
Issue Date: 2022
Publisher: Brunel University London
Abstract: Capsules (vector-valued neurons) have recently become a more active area of research in neural networks. However, existing formulations have several drawbacks including the large number of trainable parameters that they require as well as the reliance on routing mechanisms between layers of capsules. The primary aim of this project is to demonstrate the benefits of a new formulation of capsules called Homogeneous Vector Capsules (HVCs) that overcome these drawbacks. Using HVCs, new state-of-the-art accuracies for the MNIST dataset are established for multiple individual models as well as multiple ensembles. This work additionally presents a dataset consisting of high-resolution images of 13 micro-PCBs captured in various rotations and perspectives relative to the camera, with each sample labeled for PCB type, rotation category, and perspective categories. Experiments performed and elucidated in this work examine classification accuracy of rotations and perspectives that were not trained on as well as the ability to artificially generate missing rotations and perspectives during training. The results of these experiments include showing that using HVCs is superior to using fully connected layers. This work also showed that certain training samples are more informative of class membership than others. These samples can be identified prior to training by analyzing their position in reduced dimensional space relative to the classes’ centroids in that space. And a definition and calculation both for class density and dataset completeness based on the distribution of data in the reduced dimensional space has been put forth. Experimentation using the dataset completeness calculation shows that those datasets that meet a certain completeness threshold can be trained on a subset of the total dataset, based on each class’s density, while improving upon or maintaining validation accuracy.
Description: This thesis was submitted for the degree of Doctor of Philosophy and awarded by Brunel University London
Appears in Collections:Electronic and Computer Engineering
Dept of Electronic and Computer Engineering Theses

Files in This Item:
File Description SizeFormat 
FulltextThesis.pdf11.39 MBAdobe PDFView/Open

Items in BURA are protected by copyright, with all rights reserved, unless otherwise indicated.